This study explores the enhancement of channel state information (CSI) feedback in wireless communication systems by applying artificial intelligence/machine learning (AI/ML). The traditional joint training of an AI encoder at the user equipment (UE) and an AI decoder at the network (NW) side presents several challenges. Jointly trained models require sharing proprietary information, increase vulnerability to adversarial attacks, and are less suited for multi-user or multi-base station scenarios. In response to these challenges, the approach of training model entities independently has garnered interest, centring on two primary methods: UE-first separate training and NW-first sepa-rate training. Empirical findings from Release 18 AI/ML for Air Interface studies indicate that the UE-first approach yields better performance. However, this method limits the network's flexibility to accommodate distinct NW -side decoder model configurations tailored to various cells and sites. In response, this paper intro-duces a novel enhancement to the conventional NW-first separate training strategy to achieve performance gains, particularly at low quantizer resolution. Our results confirm that both the improved NW-first and UE-first strategies deliver comparable performance, both nearing joint training benchmarks.